Add support for acquire, release, and acq_rel memory ordering in PNaCl

test/Transforms/NaCl/resolve-pnacl-intrinsics.ll

Index: test/Transforms/NaCl/resolve-pnacl-intrinsics.ll
diff --git a/test/Transforms/NaCl/resolve-pnacl-intrinsics.ll b/test/Transforms/NaCl/resolve-pnacl-intrinsics.ll
index 1f6a1a860dae420f904076a4712e9fbda4b9e2a7..6cabefcc39b450e69e0d699549a16ec699486c21 100644
--- a/test/Transforms/NaCl/resolve-pnacl-intrinsics.ll
+++ b/test/Transforms/NaCl/resolve-pnacl-intrinsics.ll
@@ -1,4 +1,5 @@
-; RUN: opt < %s -resolve-pnacl-intrinsics -S | FileCheck %s -check-prefix=CLEANED
+; RUN: opt < %s -resolve-pnacl-intrinsics -S | FileCheck %s \
+; RUN:   -check-prefix=CLEANED
 ; RUN: opt < %s -resolve-pnacl-intrinsics -S | FileCheck %s
 
 ; CLEANED-NOT: call i32 @llvm.nacl.setjmp
@@ -53,53 +54,79 @@ define void @call_longjmp(i8* %arg, i32 %num) {
 
 ; atomics.
 
-; CHECK: @test_fetch_and_add_i32
+; CHECK-LABEL: @test_atomic_acquire
+define i32 @test_atomic_acquire(i32* %ptr) {
+  ; CHECK: %1 = load atomic i32* %ptr acquire, align 4
+  %1 = call i32 @llvm.nacl.atomic.load.i32(i32* %ptr, i32 3)
+  ret i32 %1
+}
+
+; CHECK-LABEL: @test_atomic_release
+define void @test_atomic_release(i32* %ptr, i32 %value) {
+  ; CHECK: store atomic i32 %value, i32* %ptr release, align 4
+  call void @llvm.nacl.atomic.store.i32(i32 %value, i32* %ptr, i32 4)
+  ret void
+}
+
+; CHECK-LABEL: @test_atomic_acquire_release
+define i32 @test_atomic_acquire_release(i32* %ptr, i32 %value) {
+  ; CHECK: %1 = atomicrmw add i32* %ptr, i32 %value acq_rel
+  %1 = call i32 @llvm.nacl.atomic.rmw.i32(i32 1, i32* %ptr, i32 %value, i32 5)
+  ret i32 %1
+}
+
+; CHECK-LABEL: @test_fetch_and_add_i32
 define i32 @test_fetch_and_add_i32(i32* %ptr, i32 %value) {
   ; CHECK: %1 = atomicrmw add i32* %ptr, i32 %value seq_cst
   %1 = call i32 @llvm.nacl.atomic.rmw.i32(i32 1, i32* %ptr, i32 %value, i32 6)
   ret i32 %1
 }
 
-; CHECK: @test_fetch_and_sub_i32
+; CHECK-LABEL: @test_fetch_and_sub_i32
 define i32 @test_fetch_and_sub_i32(i32* %ptr, i32 %value) {
   ; CHECK: %1 = atomicrmw sub i32* %ptr, i32 %value seq_cst
   %1 = call i32 @llvm.nacl.atomic.rmw.i32(i32 2, i32* %ptr, i32 %value, i32 6)
   ret i32 %1
 }
 
-; CHECK: @test_fetch_and_or_i32
+; CHECK-LABEL: @test_fetch_and_or_i32
 define i32 @test_fetch_and_or_i32(i32* %ptr, i32 %value) {
   ; CHECK: %1 = atomicrmw or i32* %ptr, i32 %value seq_cst
   %1 = call i32 @llvm.nacl.atomic.rmw.i32(i32 3, i32* %ptr, i32 %value, i32 6)
   ret i32 %1
 }
 
-; CHECK: @test_fetch_and_and_i32
+; CHECK-LABEL: @test_fetch_and_and_i32
 define i32 @test_fetch_and_and_i32(i32* %ptr, i32 %value) {
   ; CHECK: %1 = atomicrmw and i32* %ptr, i32 %value seq_cst
   %1 = call i32 @llvm.nacl.atomic.rmw.i32(i32 4, i32* %ptr, i32 %value, i32 6)
   ret i32 %1
 }
 
-; CHECK: @test_fetch_and_xor_i32
+; CHECK-LABEL: @test_fetch_and_xor_i32
 define i32 @test_fetch_and_xor_i32(i32* %ptr, i32 %value) {
   ; CHECK: %1 = atomicrmw xor i32* %ptr, i32 %value seq_cst
   %1 = call i32 @llvm.nacl.atomic.rmw.i32(i32 5, i32* %ptr, i32 %value, i32 6)
   ret i32 %1
 }
 
-; CHECK: @test_val_compare_and_swap_i32
+; Test different compare-and-swap patterns that commonly occur and are a bit
+; tricky because the PNaCl intrinsic only returns the value whereas the LLVM
+; intrinsic also returns the success flag (equivalent to comparing the oldval
+; with what was just loaded).
+
+; CHECK-LABEL: @test_val_compare_and_swap_i32
 define i32 @test_val_compare_and_swap_i32(i32* %ptr, i32 %oldval, i32 %newval) {
-  ; CHECK: %1 = cmpxchg i32* %ptr, i32 %oldval, i32 %newval seq_cst
+  ; CHECK: %1 = cmpxchg i32* %ptr, i32 %oldval, i32 %newval seq_cst seq_cst
   ; CHECK-NEXT: %2 = extractvalue { i32, i1 } %1, 0
   ; CHECK-NEXT: ret i32 %2
   %1 = call i32 @llvm.nacl.atomic.cmpxchg.i32(i32* %ptr, i32 %oldval, i32 %newval, i32 6, i32 6)
   ret i32 %1
 }
 
-; CHECK: @test_val_compare_and_swap_i32_new
+; CHECK-LABEL: @test_val_compare_and_swap_i32_new
 define i32 @test_val_compare_and_swap_i32_new(i32* %ptr, i32 %oldval, i32 %newval) {
-  ; CHECK: %1 = cmpxchg i32* %ptr, i32 %oldval, i32 %newval seq_cst
+  ; CHECK: %1 = cmpxchg i32* %ptr, i32 %oldval, i32 %newval seq_cst seq_cst
   ; CHECK-NEXT: %res2 = extractvalue { i32, i1 } %1, 0
   ; CHECK-NEXT: ret i32 %res2
   %res = call i32 @llvm.nacl.atomic.cmpxchg.i32(i32* %ptr, i32 %oldval, i32 %newval, i32 6, i32 6)
@@ -110,9 +137,9 @@ define i32 @test_val_compare_and_swap_i32_new(i32* %ptr, i32 %oldval, i32 %newva
   ret i32 %val
 }
 
-; CHECK: @test_bool_compare_and_swap_i32
+; CHECK-LABEL: @test_bool_compare_and_swap_i32
 define i1 @test_bool_compare_and_swap_i32(i32* %ptr, i32 %oldval, i32 %newval) {
-  ; CHECK: %1 = cmpxchg i32* %ptr, i32 %oldval, i32 %newval seq_cst
+  ; CHECK: %1 = cmpxchg i32* %ptr, i32 %oldval, i32 %newval seq_cst seq_cst
   ; CHECK-NEXT: %success = extractvalue { i32, i1 } %1, 1
   ; CHECK-NEXT: ret i1 %success
   %1 = call i32 @llvm.nacl.atomic.cmpxchg.i32(i32* %ptr, i32 %oldval, i32 %newval, i32 6, i32 6)
@@ -120,9 +147,9 @@ define i1 @test_bool_compare_and_swap_i32(i32* %ptr, i32 %oldval, i32 %newval) {
   ret i1 %2
 }
 
-; CHECK: @test_bool_compare_and_swap_i32_new
+; CHECK-LABEL: @test_bool_compare_and_swap_i32_new
 define i1 @test_bool_compare_and_swap_i32_new(i32* %ptr, i32 %oldval, i32 %newval) {
-  ; CHECK: %1 = cmpxchg i32* %ptr, i32 %oldval, i32 %newval seq_cst
+  ; CHECK: %1 = cmpxchg i32* %ptr, i32 %oldval, i32 %newval seq_cst seq_cst
   ; CHECK-NEXT: %suc = extractvalue { i32, i1 } %1, 1
   ; CHECK-NEXT: ret i1 %suc
   %res = call i32 @llvm.nacl.atomic.cmpxchg.i32(i32* %ptr, i32 %oldval, i32 %newval, i32 6, i32 6)
@@ -133,9 +160,9 @@ define i1 @test_bool_compare_and_swap_i32_new(i32* %ptr, i32 %oldval, i32 %newva
   ret i1 %suc
 }
 
-; CHECK: @test_bool_compare_and_swap_i32_reordered
+; CHECK-LABEL: @test_bool_compare_and_swap_i32_reordered
 define i1 @test_bool_compare_and_swap_i32_reordered(i32* %ptr, i32 %oldval, i32 %newval) {
-  ; CHECK: %1 = cmpxchg i32* %ptr, i32 %oldval, i32 %newval seq_cst
+  ; CHECK: %1 = cmpxchg i32* %ptr, i32 %oldval, i32 %newval seq_cst seq_cst
   ; CHECK-NEXT: %success = extractvalue { i32, i1 } %1, 1
   ; CHECK-NEXT: ret i1 %success
   %1 = call i32 @llvm.nacl.atomic.cmpxchg.i32(i32* %ptr, i32 %oldval, i32 %newval, i32 6, i32 6)
@@ -143,9 +170,9 @@ define i1 @test_bool_compare_and_swap_i32_reordered(i32* %ptr, i32 %oldval, i32
   ret i1 %2
 }
 
-; CHECK: @test_struct_compare_and_swap_i32
+; CHECK-LABEL: @test_struct_compare_and_swap_i32
 define { i32, i1 } @test_struct_compare_and_swap_i32(i32* %ptr, i32 %oldval, i32 %newval) {
-  ; CHECK: %1 = cmpxchg i32* %ptr, i32 %oldval, i32 %newval seq_cst
+  ; CHECK: %1 = cmpxchg i32* %ptr, i32 %oldval, i32 %newval seq_cst seq_cst
   ; CHECK-NEXT: ret { i32, i1 } %1
   %1 = call i32 @llvm.nacl.atomic.cmpxchg.i32(i32* %ptr, i32 %oldval, i32 %newval, i32 6, i32 6)
   %2 = icmp eq i32 %1, %oldval
@@ -154,14 +181,37 @@ define { i32, i1 } @test_struct_compare_and_swap_i32(i32* %ptr, i32 %oldval, i32
   ret { i32, i1 } %4
 }
 
-; CHECK: @test_c11_fence
+; Test all allowed cmpxchg success/failure memory orderings.
+
+; CHECK-LABEL: @test_cmpxchg_seqcst_seqcst
+define i32 @test_cmpxchg_seqcst_seqcst(i32* %ptr, i32 %oldval, i32 %newval) {
+  ; CHECK: %1 = cmpxchg i32* %ptr, i32 %oldval, i32 %newval seq_cst seq_cst
+  %1 = call i32 @llvm.nacl.atomic.cmpxchg.i32(i32* %ptr, i32 %oldval, i32 %newval, i32 6, i32 6)
+  ret i32 %1
+}
+
+; CHECK-LABEL: @test_cmpxchg_seqcst_acquire
+define i32 @test_cmpxchg_seqcst_acquire(i32* %ptr, i32 %oldval, i32 %newval) {
+  ; CHECK: %1 = cmpxchg i32* %ptr, i32 %oldval, i32 %newval seq_cst acquire
+  %1 = call i32 @llvm.nacl.atomic.cmpxchg.i32(i32* %ptr, i32 %oldval, i32 %newval, i32 6, i32 3)
+  ret i32 %1
+}
+
+; CHECK-LABEL: @test_cmpxchg_acquire_acquire
+define i32 @test_cmpxchg_acquire_acquire(i32* %ptr, i32 %oldval, i32 %newval) {
+  ; CHECK: %1 = cmpxchg i32* %ptr, i32 %oldval, i32 %newval acquire acquire
+  %1 = call i32 @llvm.nacl.atomic.cmpxchg.i32(i32* %ptr, i32 %oldval, i32 %newval, i32 3, i32 3)
+  ret i32 %1
+}
+
+; CHECK-LABEL: @test_c11_fence
 define void @test_c11_fence() {
   ; CHECK: fence seq_cst
   call void @llvm.nacl.atomic.fence(i32 6)
   ret void
 }
 
-; CHECK: @test_synchronize
+; CHECK-LABEL: @test_synchronize
 define void @test_synchronize() {
   ; CHECK: call void asm sideeffect "", "~{memory}"()
   ; CHECK: fence seq_cst
@@ -170,14 +220,14 @@ define void @test_synchronize() {
   ret void
 }
 
-; CHECK: @test_is_lock_free_1
+; CHECK-LABEL: @test_is_lock_free_1
 define i1 @test_is_lock_free_1(i8* %ptr) {
   ; CHECK: ret i1 {{true|false}}
   %res = call i1 @llvm.nacl.atomic.is.lock.free(i32 1, i8* %ptr)
   ret i1 %res
 }
 
-; CHECK: @test_is_lock_free_2
+; CHECK-LABEL: @test_is_lock_free_2
 define i1 @test_is_lock_free_2(i16* %ptr) {
   ; CHECK: ret i1 {{true|false}}
   %ptr2 = bitcast i16* %ptr to i8*
@@ -185,7 +235,7 @@ define i1 @test_is_lock_free_2(i16* %ptr) {
   ret i1 %res
 }
 
-; CHECK: @test_is_lock_free_4
+; CHECK-LABEL: @test_is_lock_free_4
 define i1 @test_is_lock_free_4(i32* %ptr) {
   ; CHECK: ret i1 {{true|false}}
   %ptr2 = bitcast i32* %ptr to i8*
@@ -193,7 +243,7 @@ define i1 @test_is_lock_free_4(i32* %ptr) {
   ret i1 %res
 }
 
-; CHECK: @test_is_lock_free_8
+; CHECK-LABEL: @test_is_lock_free_8
 define i1 @test_is_lock_free_8(i64* %ptr) {
   ; CHECK: ret i1 {{true|false}}
   %ptr2 = bitcast i64* %ptr to i8*
@@ -201,14 +251,14 @@ define i1 @test_is_lock_free_8(i64* %ptr) {
   ret i1 %res
 }
 
-; CHECK: @test_lock_test_and_set_i32
+; CHECK-LABEL: @test_lock_test_and_set_i32
 define i32 @test_lock_test_and_set_i32(i32* %ptr, i32 %value) {
   ; CHECK: %1 = atomicrmw xchg i32* %ptr, i32 %value seq_cst
   %1 = call i32 @llvm.nacl.atomic.rmw.i32(i32 6, i32* %ptr, i32 %value, i32 6)
   ret i32 %1
 }
 
-; CHECK: @test_lock_release_i32
+; CHECK-LABEL: @test_lock_release_i32
 define void @test_lock_release_i32(i32* %ptr) {
   ; Note that the 'release' was changed to a 'seq_cst'.
   ; CHECK: store atomic i32 0, i32* %ptr seq_cst, align 4
@@ -216,56 +266,56 @@ define void @test_lock_release_i32(i32* %ptr) {
   ret void
 }
 
-; CHECK: @test_atomic_load_i8
+; CHECK-LABEL: @test_atomic_load_i8
 define zeroext i8 @test_atomic_load_i8(i8* %ptr) {
   ; CHECK: %1 = load atomic i8* %ptr seq_cst, align 1
   %1 = call i8 @llvm.nacl.atomic.load.i8(i8* %ptr, i32 6)
   ret i8 %1
 }
 
-; CHECK: @test_atomic_store_i8
+; CHECK-LABEL: @test_atomic_store_i8
 define void @test_atomic_store_i8(i8* %ptr, i8 zeroext %value) {
   ; CHECK: store atomic i8 %value, i8* %ptr seq_cst, align 1
   call void @llvm.nacl.atomic.store.i8(i8 %value, i8* %ptr, i32 6)
   ret void
 }
 
-; CHECK: @test_atomic_load_i16
+; CHECK-LABEL: @test_atomic_load_i16
 define zeroext i16 @test_atomic_load_i16(i16* %ptr) {
   ; CHECK: %1 = load atomic i16* %ptr seq_cst, align 2
   %1 = call i16 @llvm.nacl.atomic.load.i16(i16* %ptr, i32 6)
   ret i16 %1
 }
 
-; CHECK: @test_atomic_store_i16
+; CHECK-LABEL: @test_atomic_store_i16
 define void @test_atomic_store_i16(i16* %ptr, i16 zeroext %value) {
   ; CHECK: store atomic i16 %value, i16* %ptr seq_cst, align 2
   call void @llvm.nacl.atomic.store.i16(i16 %value, i16* %ptr, i32 6)
   ret void
 }
 
-; CHECK: @test_atomic_load_i32
+; CHECK-LABEL: @test_atomic_load_i32
 define i32 @test_atomic_load_i32(i32* %ptr) {
   ; CHECK: %1 = load atomic i32* %ptr seq_cst, align 4
   %1 = call i32 @llvm.nacl.atomic.load.i32(i32* %ptr, i32 6)
   ret i32 %1
 }
 
-; CHECK: @test_atomic_store_i32
+; CHECK-LABEL: @test_atomic_store_i32
 define void @test_atomic_store_i32(i32* %ptr, i32 %value) {
   ; CHECK: store atomic i32 %value, i32* %ptr seq_cst, align 4
   call void @llvm.nacl.atomic.store.i32(i32 %value, i32* %ptr, i32 6)
   ret void
 }
 
-; CHECK: @test_atomic_load_i64
+; CHECK-LABEL: @test_atomic_load_i64
 define i64 @test_atomic_load_i64(i64* %ptr) {
   ; CHECK: %1 = load atomic i64* %ptr seq_cst, align 8
   %1 = call i64 @llvm.nacl.atomic.load.i64(i64* %ptr, i32 6)
   ret i64 %1
 }
 
-; CHECK: @test_atomic_store_i64
+; CHECK-LABEL: @test_atomic_store_i64
 define void @test_atomic_store_i64(i64* %ptr, i64 %value) {
   ; CHECK: store atomic i64 %value, i64* %ptr seq_cst, align 8
   call void @llvm.nacl.atomic.store.i64(i64 %value, i64* %ptr, i32 6)